design-cute-dsl-kernel

Structured workflow skill for learning CuTe Python DSL by reading, importing, profiling, and extracting reusable patterns from CUTLASS Blackwell example kernels for GPU programming.

Compatibility router skill for a shared optimization knowledge base and language-specific optimization catalogs, dispatching between CuTe DSL and other implementation languages for GPU kernel optimization.

API reference and implementation patterns for NVIDIA's CuTe Python DSL, covering the two-level host/device kernel model, TMA pipelines, barrier choreography, shared memory management, and the three-stage JIT compilation pipeline for GPU kernel programming.

A skill providing language-specific guidance for cuTile Python DSL kernel implementation, covering the CtKernel runtime wrapper, suitability gate for deciding when to stay in cuTile vs. switching to CuTe DSL, and cuTile-specific pitfalls for GPU kernel optimization.

Optimized Multi-Latent Attention (MLA) kernels for DeepSeek-V3/R1 and Kimi-V2 inference. Bridges the performance gap between FlashMLA (decode) and FlashAttention (prefill) for speculative decoding regimes, with profiling tools and a double-versioned kernel structure for easy comparison.

Provides cost models and regime analysis for Multi-Latent Attention (MLA), comparing FlashMLA, FlashAttention, and MLAvar6+ kernels across decode, speculative, and prefill regimes to guide kernel selection and optimization.

A GPU kernel profiling workflow skill supporting four modes (annotation, event, NCU, NSYS) with device-aware peak calculations, bottleneck identification rules, and roofline analysis for optimizing kernel performance.

A skill providing GPU kernel optimization guidance specific to the CuTe Python DSL, layered on top of shared optimization and anti-pattern catalogs. It helps determine whether findings are shared or CuTe DSL-specific and records implementation notes.

A skill providing GPU kernel optimization guidance specific to the cuTile Python DSL, layered on top of shared optimization and anti-pattern catalogs. It covers cuTile-specific topics like num_ctas, occupancy, math semantics, and code-generation sensitivity.

A skill defining the full GPU kernel optimization workflow, including sub-agent launch templates, agent communication contracts, default configurations for speculative decoding, and tuning strategies. It coordinates profiler and designer agents through iterative optimization cycles.